acceptodds
Under review as a conference paper at ICLR 2027

From Latent Representations to Causal Health States: Degradation-Aware Geometry for Remaining Useful Life Prediction

Abstract

Remaining Useful Life (RUL) prediction is commonly formulated as direct regression from condition-monitoring signals to a scalar lifetime estimate. While increasingly powerful models achieve strong predictive performance, this formulation does not explicitly ensure that their learned representations reflect the underlying degradation process. We introduce Health-Aware Manifold Learning (HAML), a new representation-backend-portable framework that constructs an explicit degradation geometry over latent representations before RUL prediction. HAML first encodes causal sensor windows into a low-dimensional representation without RUL or lifecycle-progress supervision. Using training data only, lifecycle progress guides the construction of a sparse degradation-aware landmark graph that penalises cross-stage shortcuts through a progress-aware graph metric. Distributed healthy and failure regions define a graph-geodesic health coordinate, which a persistent level-velocity filter transforms into a bounded, non-decreasing Health Index (HI) using only observations available up to the current cycle. The resulting health dynamics and complementary manifold geometry support downstream RUL estimation. Experiments on all four C-MAPSS aircraft-engine subsets and the PHM12 bearing dataset demonstrate competitive RUL prediction, meaningful lifecycle alignment, and favourable late-life error behaviour, while controlled C-MAPSS sensor perturbations demonstrate robustness to test-time noise. Controlled ablations clarify the roles of individual components, while alternative representation backends demonstrate HAML's portability beyond UMAP. These results show that explicitly modelling degradation geometry and its temporal evolution offers a principled alternative to unconstrained sensor-to-RUL regression.

Primary area:
learning on time series and dynamical systems
Keywords:
Remaining Useful Life prediction, Prognostics and Health Management, manifold learning, health index, graph-geodesic learning, degradation modelling, causal filtering, predictive maintenance

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_oSrfOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Health-Aware Manifold Learning (HAML) for remaining useful life (RUL) prediction. It first fits a representation without lifecycle labels, then introduces lifecycle supervision through a progress-aware landmark graph and distributed healthy/failure reference regions. Relative graph-geodesic distances define a health coordinate, which a persistent level–velocity filter converts into a causal, bounded, non-decreasing health index (HI). A downstream decoder combines HI history, temporal features, and complementary geometric information. Experiments on C-MAPSS and PHM12 assess predictive performance, health-trajectory diagnostics, robustness, and representation-backend portability.

Critical strengths

1.A useful and explicit formulation of health-state construction. HAML separates representation learning, supervised health geometry, causal trajectory construction, and RUL decoding. Distributed endpoint references accommodate heterogeneous degradation states, while the filter imposes clearly specified online trajectory constraints. The paper appropriately limits its theoretical claims and explicitly acknowledges that these constraints do not guarantee predictive optimality or physical validity.

2.Substantial experimental coverage and informative controls. The evaluation includes two benchmark families, ten-seed reporting, noise perturbations, eight ablation groups, and multiple representation backends. Figure G1 compares the original descriptors, UMAP coordinates, and HAML features across several decoders; Table G1 additionally supplies descriptor-matched neural baselines. These are valuable controls. The authors also report unfavorable results, including cases where direct descriptor-based prediction is stronger and where removing filtering improves accuracy.

Critical weaknesses

1.The population and implementation used for the HI diagnostics are insufficiently specified, and some reported results are difficult to reconcile with their definitions. Appendix D states that PHM12 Condition 3 has only one test bearing, whereas Eq. D7 defines prognosability using the standard deviation of failure-endpoint HI values across assets. Table G5 nevertheless reports 0.996 ± 0.003 for this condition. If the metric is computed separately for each seed on that single held-out bearing, the population standard deviation is zero, giving prognosability 1 when the denominator is nonzero; the sample standard deviation would be undefined. The evaluated bearing population and aggregation across assets and seeds therefore need to be stated. In addition, Eq. D5 uses signed increments: under the usual convention sgn(0) = 0, a non-decreasing trajectory containing plateaus need not receive a score of 1, contrary to the accompanying construction-based guarantee. These observations do not establish an implementation error, but they prevent a clear interpretation of evidence central to the paper’s health-state claims.

2.The additional value of the constrained health state is not sufficiently separated from imposed temporal structure. In Table 2, removing filtering (A5) improves the reported mean RMSE, S-score, and P95 Late Error on all four C-MAPSS subsets. For example, RMSE changes from 13.13 to 12.77 on FD001 and from 14.31 to 13.73 on FD003. The authors explicitly acknowledge this accuracy–constraint trade-off, which is reasonable. However, the compensating health-state benefit needs stronger validation. In Eq. D6, normalized lifecycle progress is an affine transformation of time within each asset, so the reported per-asset Pearson correlation is also correlation with time. This measures temporal alignment but does not by itself establish additional sensor-derived degradation information. Furthermore, replacing the filter with cumulative maximum in A6 changes several mechanisms simultaneously, so it does not cleanly isolate persistent velocity memory. The full-pipeline noise experiments likewise do not isolate the filter’s contribution. An age-only diagnostic control and a comparison retaining the smoothing and projection rules while removing persistent velocity would help establish the specific benefit.

Questions for authors

1.Which exact assets and trajectory segments enter the HI diagnostics in Tables G2 and G5, and are these assets excluded from representation fitting, graph construction, and reference selection? How are statistics aggregated across assets and seeds, particularly for PHM12 Condition 3?

2.How are zero increments handled in the implementation of Eq. D5? Does the reported metric implement the displayed sign-based expression or a non-negative-increment proportion?

3.What evidence shows that the persistent level–velocity mechanism provides useful health-state information beyond age alignment and generic monotone smoothing? In particular, is its claimed benefit retained when smoothing and projection are held fixed, or can the existing noise evaluation distinguish the full filter from A5 and a matched simpler filter?

6 comments
AuthorsOpenReview ↗

Response to reviewer oSrf (part 1)

We thank the reviewer for their constructive feedback on our methodology, particularly on HI evaluation and the distinction between imposed temporal constraints and sensor-derived degradation information. To address these concerns, we conducted additional comparisons against age-only baselines and matched velocity-filter ablations (Tables R2–R3). These results provide further evidence that the learned health representations capture sensor-derived information beyond elapsed age, and that velocity persistence improves HI diagnostics across the training lifecycle. We address each point below.

1. HI evaluation population and aggregation (Q1, W1). Tables G2/G5 and Fig. 3/Table G6 use complete run-to-failure training trajectories: C-MAPSS FD001–FD004 engines from their first valid 50-cycle-window endpoint to failure, and PHM12 bearings 1_1/1_2, 2_1/2_2 and 3_1/3_2. These assets also enter representation fitting, graph construction, and healthy/failure reference selection. Thus, the diagnostics describe in-sample lifecycle structure and constraint compliance, not held-out HI generalisation. We will correct the misleading “held-out lifecycle progress” wording in Section 4.4. Held-out RUL evaluation remains separate and uses no future test observations or failure endpoints. This distinction follows the endpoint-dependent definition of prognosability in Coble and Hines [1]; Dama and Sinoquet [2] provide a C-MAPSS precedent for health-state construction on complete training histories with separate RUL evaluation on test histories.

For PHM12 Condition 3, Eq. D7 is evaluated within each seed on training bearings 3_1 and 3_2, not the single test bearing 3_3. The numerator is the population SD of their terminal HI values; the denominator is their mean absolute start-to-end HI change. The result 0.996 ± 0.003 is the mean ± SD over ten seed-level scores, not uncertainty over one test bearing. We will state the asset sets and aggregation rules explicitly, and acknowledge that two training bearings provide limited evidence of population-level prognosability.

References

[1] Coble, J., & Hines, J. W. (2009). Identifying optimal prognostic parameters from data: A genetic algorithms approach. Annual Conference of the PHM Society, 1(1).

[2] Dama, F., & Sinoquet, C. (2023). Partially Hidden Markov Chain Multivariate Linear Autoregressive model: Inference and forecasting—application to machine health prognostics. Machine Learning, 112, 45–97.

2. Beyond age alignment (W2). HI-progress correlation alone does not demonstrate sensor-derived degradation information. We therefore compared HAML with linear (Age-L) HI(t)=min⁡(max⁡(t−50T−50,0),1)HI(t)=\min\left(\max\left(\frac{t-50}{T-50},0\right),1\right) and isotonic (Age-NL)HI(t)=fiso(t)HI(t)=f_{\mathrm{iso}}(t) age-only controls using five-fold, engine-disjoint evaluation over ten split seeds across FD001-FD004. Preprocessing, UMAP, graph fitting, and filter calibration used calibration engines only; HI diagnostics used held-out complete trajectories. HAML achieved higher healthy–failure separation than Age-L (0.887–0.930 vs 0.843–0.881) and prognosability than Age-NL (0.984–0.997 vs 0.897–0.920), although Age-L had higher progress correlation (0.976–0.990 vs 0.896–0.925). On official test engines, full HAML outperformed Age-NL in RUL prediction (FD003 RMSE 14.31 vs 31.84). These results support sensor-informed information beyond age, although terminal HI saturation (94–98%) and unmatched decoder capacity limit attribution to HI alone (Table R2).

Table R2. Engine-disjoint evaluation of HAML versus age-only controls: held-out HI diagnostics and test-set RUL prediction.

DatasetPearson (D6) ↑ Age-L / HAMLPrognosability (D7) ↑ Age-NL / HAMLSeparation (D9) ↑ Age-L / HAMLRMSE ↓ Age-NL / HAMLNASA S-score ↓ Age-NL / HAML
FD0010.990 / 0.9250.916 / 0.9970.881 / 0.93032.07 / 13.133,931 / 274.3
FD0020.990 / 0.9180.920 / 0.9930.877 / 0.91229.56 / 14.816,817 / 884.3
FD0030.976 / 0.9020.919 / 0.9950.864 / 0.91331.84 / 14.313,827 / 1,069.1
FD0040.985 / 0.8960.897 / 0.9840.843 / 0.88737.63 / 15.7921,805 / 1,050.1

HI diagnostics are evaluated on held-out complete run-to-failure engines using five-fold, engine-disjoint validation repeated over ten splits. Age-L denotes linear age, and Age-NL denotes isotonic nonlinear age. The displayed comparators are Age-L for Pearson correlation and separation, and Age-NL for prognosability and RUL prediction. RUL results use official held-out test engines; no age-only predictor uses an individual test engine's failure time. The RUL comparison is not decoder-capacity-matched. HAML terminal HI saturation may inflate prognosability.

AuthorsOpenReview ↗

Response to reviewer oSrf (part 2)

3. Eq. D5 and zero increments (Q2, W1). We agree: with sgn⁡(0)=0\operatorname{sgn}(0)=0, the displayed sign-based formula is inconsistent with counting plateaus as non-decreasing steps. The metric described by our implementation is the fraction of adjacent pairs satisfying Δhi,t≥−10−8\Delta h_{i,t} \geq -10^{-8}, pooled across eligible training-asset pairs within each seed and reported as mean ± SD across ten seeds. We have verified that the reported values were computed using the indicator-based definition and remain unchanged. We will replace Eq. D5 accordingly and clarify the numerical tolerance and aggregation procedure. A monotonicity score of 1.000 verifies the imposed projection (Eq. 13); it is not evidence of physical degradation validity.

4. Velocity persistence versus smoothing (Q3, W2). Removing filtering (A5) improves clean-data prediction but sacrifices monotonicity; cumulative maximum (A6) alters multiple mechanisms and cannot isolate velocity persistence. We therefore reran matched A0, V1 (ρ=0\rho = 0; no carried velocity) and V2 (β=0\beta = 0, v0=0v_0 = 0; no velocity contribution) over the same ten seeds, preserving graph features, smoothing/projection rules and 49-dimensional decoder inputs. All variants remain monotone.

A0 yields higher training-lifecycle prognosability (0.984–0.999 vs. 0.965–0.984) and healthy-failure separation (0.809–0.914 versus 0.767–0.870). Exact paired sign-flip tests with Holm correction across 24 HI comparisons (3 metrics × 2 contrasts × 4 datasets) yield a minimum Holm-adjusted p-value of 0.0469. Predictive effects are smaller: none of the 24 paired RUL comparisons remains significant after Holm correction (minimum adjusted p = 0.281). Some late-error outcomes favour the ablations (Table R3). Thus, the full filter improves training-lifecycle HI diagnostics but does not demonstrate statistically significant RUL superiority over the matched controls. The existing noise experiment does not isolate the filter's contribution. Detailed statistical results will be provided in the appendix. The existing noise experiment does not isolate the filter (Table R3).

Table R3. Velocity-component ablations: training-lifecycle HI diagnostics and held-out RUL prediction.

DatasetVariantHI–progress r ↑Prognosability ↑Healthy–failure separation ↑RMSE ↓NASA S-score ↓P95 late error ↓
FD001A00.903 ± 0.0070.998 ± 0.0040.914 ± 0.01013.11 ± 0.50270.9 ± 41.524.40 ± 3.07
V10.901 ± 0.0060.984 ± 0.0050.870 ± 0.00913.23 ± 0.64277.3 ± 49.625.13 ± 3.73
V20.901 ± 0.0060.984 ± 0.0050.870 ± 0.00913.24 ± 0.68277.3 ± 51.025.08 ± 3.52
FD002A00.893 ± 0.0070.998 ± 0.0020.897 ± 0.00914.86 ± 0.63891.3 ± 157.729.92 ± 2.36
V10.892 ± 0.0070.980 ± 0.0050.844 ± 0.01015.04 ± 0.57937.2 ± 147.630.07 ± 1.99
V20.892 ± 0.0070.980 ± 0.0050.844 ± 0.01015.01 ± 0.52924.9 ± 128.830.03 ± 2.14
FD003A00.872 ± 0.0090.999 ± 0.0020.890 ± 0.00714.39 ± 0.70922.3 ± 311.233.13 ± 1.86
V10.870 ± 0.0080.983 ± 0.0050.856 ± 0.00814.41 ± 0.58936.4 ± 266.432.61 ± 1.56
V20.870 ± 0.0080.983 ± 0.0050.856 ± 0.00814.41 ± 0.59949.0 ± 277.532.61 ± 1.58
FD004A00.859 ± 0.0130.984 ± 0.0170.809 ± 0.08515.85 ± 0.341090.3 ± 78.028.05 ± 2.78
V10.858 ± 0.0130.965 ± 0.0120.767 ± 0.08115.85 ± 0.381087.0 ± 103.627.75 ± 1.84
V20.858 ± 0.0130.965 ± 0.0120.767 ± 0.08115.88 ± 0.431095.6 ± 114.028.02 ± 1.83

A0 = full HAML rerun for this experiment; V1 = no velocity persistence (ρ=0\rho = 0); V2 = no velocity contribution (β=0\beta = 0, v0=0v_0 = 0). Results are mean ± SD over ten matched seeds. HI diagnostics are computed on complete training trajectories, whereas RUL metrics are evaluated on official held-out test assets. All configurations retain smoothing and projection and achieve monotonicity of 1.000. A0 was independently rerun alongside V1/V2 under the matched experimental protocol. The resulting A0 values are reported separately from the original manuscript results to ensure within-experiment comparability. Lower RMSE, NASA S-score, and P95 Late Error are better; higher HI correlation, prognosability, and separation are better.

Reviewer_oSrfOpenReview ↗

Follow-up on HI Evaluation and Remaining Concerns

Thank you for the clarifications and additional experiments. The released code supports your explanation of the implemented monotonicity metric and the use of training trajectories for the original HI diagnostics. I appreciate the engine-disjoint age-only comparisons and the matched velocity ablations, which address important parts of my concerns.

However, the remaining evidence is not yet sufficient to establish the additional value of the proposed health-state construction. In Table R2, terminal HI saturation limits the interpretation of prognosability, while the unmatched decoder capacity prevents attributing the RUL advantage specifically to the HI. Table R3 provides a better-controlled comparison, but its HI improvements are measured on training trajectories, and no RUL comparison remains significant after Holm correction. This does not establish that the velocity component is ineffective, but its practical benefit remains insufficiently demonstrated.

One implementation detail also needs clarification: Eqs. D6 and D9 specify per-engine calculations followed by averaging, whereas hi_metrics in the released metrics.py computes correlation and separation over pooled observations. These procedures are generally not equivalent. Please clarify which aggregation was used for the original results and Tables R2–R3, and reconcile the definitions with the corresponding implementation.

I acknowledge the improvements in the rebuttal, but maintain my current rating for now because the central evidential concerns remain only partially resolved.

Reviewer_Qq3NOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Health-Aware Manifold Learning (HAML) for remaining useful life (RUL) prediction in engineering systems. Rather than directly regressing RUL from sensor signals, HAML constructs an explicit degradation aware geometric representation and transforms it into a bounded, non-decreasing Health Index (HI). The framework combines a temporal descriptor encoder, a progress-aware landmark graph, a graph-geodesic health coordinate, a persistent level–velocity filter, and an HI-centred RUL decoder. Experiments on all four C-MAPSS subsets and the PHM12 bearing dataset demonstrate competitive predictive performance, improved alignment between the HI and lifecycle progress, and robustness to controlled test-time sensor noise. However, HAML does not consistently outperform existing methods on every dataset and metric.

Critical strengths
  1. The paper identifies a limitation of direct sensor to RUL regression that is predictive representations are not necessarily consistent with the underlying degradation process. HAML explicitly separates representation learning from degradation-aware geometry construction, providing a coherent methodological rationale.
  2. The progress aware landmark graph penalises connections between states at different lifecycle stages, while distributed healthy and failure reference regions define a graph-geodesic health coordinate. The persistent level velocity filter then enforces bounded, non-decreasing health trajectories while retaining degradation momentum. These components are logically integrated rather than operating as isolated techniques.
  3. The ablation study is well designed and comprehensive. It examines graph geometry, progress-aware edge weighting, reference region construction, temporal edges, causal filtering, decoder information, and representation-backend selection.
  4. Figure 3 on page 8 shows that causal filtering improves Pearson and Spearman correlations with lifecycle progress across all four C-MAPSS subsets. The authors also explicitly acknowledge limitations involving irreversible degradation assumptions, representation quality, and landmark configuration.
Critical weaknesses
  1. Although HAML achieves competitive results, its performance varies substantially across datasets. On FD003, for example, its RMSE is 14.31, compared with 11.63 for GRU and 12.01 for LSTM. On FD002, TimesNet achieves an RMSE of 13.48, compared with 14.81 for HAML. These results suggest that the benefits of explicit degradation geometry are dataset-dependent. The paper would benefit from stronger statistical comparisons and clearer analysis of when the proposed framework is advantageous.
  2. The main contribution lies in introducing degradation supervision after representation fitting and combining graph-geodesic health construction with causal filtering. However, manifold based degradation representations, health index estimation and monotonic filtering are established research directions. The paper should more clearly distinguish its technical contributions from these prior approaches and demonstrate which aspects cannot be achieved by simpler combinations of existing methods.
  3. Although progress labels are not needed at inference, the training procedure still depends on run to failure trajectories and their lifecycle information. The authors should better discuss the implications for assets with incomplete failure histories, changing operating conditions, or limited degradation annotations.
  4. The causal interpretation is potentially overstated. The paper explicitly defines causality as temporal or online causality, meaning that predictions at cycle t use no future observations. It does not identify causal effects or establish a physical degradation mechanism. Consequently, the term “causal degradation states” should be interpreted carefully.
Questions for authors

1.How does HAML fundamentally differ from existing manifold based health index construction methods and monotonic health index learning? 2.HAML uses a 49-dimensional feature representation derived from the health index, its history, velocity, and latent coordinates, together with a BRF decoder. Are the baselines given comparable access to temporal information, feature engineering, and hyperparameter optimization? Could these differences explain some of the performance gains? 3.HAML performs competitively on FD001 and FD004, but several baselines outperform it on FD003, while some ablations also improve predictive metrics. What evidence demonstrates that the proposed geometric construction consistently benefits RUL prediction rather than primarily improving health index smoothness?

1 comment

We thank the reviewer for their insightful comments on HAML's dataset-dependent performance, technical contributions, and experimental fairness. We address each point below.

1. Dataset dependence and statistical significance (W1). Two-sided Welch's t-tests over ten seeds, with Holm correction (α=0.05\alpha = 0.05) across ten baselines per dataset/bearing–metric group, show that HAML significantly outperforms, underperforms, or does not differ significantly from baselines in 36/21/6336/21/63 of 120 C-MAPSS (4 subsets ×\times 3 metrics $\times$10 baselines) and 120/16/194120/16/194 of 330 PHM12 (11 bearings$\times$3 metrics$\times$10 baselines) comparisons, respectively (Detailed Welch–Holm results will be provided in the appendix). On FD004, HAML significantly improves P95 Late Error over 9/109/10 baselines, whereas 16/2116/21 C-MAPSS losses occur on FD003. These are associated with concentrated late-overestimation risk: one engine contributes 72.1$\pm$8.7% of FD003's S-score (Fig. G2). Switching to PCA reduces its S-score from 1069.1 to 322.3 (A8), highlighting backend sensitivity. Across 15 tasks, HAML achieves the joint-best average RMSE rank (4.07) and best S-score (4.10) and P95 (3.40) ranks. Its strengths are primarily risk-sensitive, while weaker RMSE on heterogeneous FD002/FD004 and the raw graph and latent features contribute 32–40% of decoder importance (Table G7), suggesting that a scalar HI may not fully capture heterogeneous degradation regimes.

2. Technical distinction (W2, Q1). We agree that each ingredient is established. The distinction is where supervision enters and what the HI is. In prior work, degradation labels shape the representation, or the HI is a regressed target or latent dimension (Yu et al.,2019; Wei et al.,2021). In HAML, the backend is fitted without lifecycle/RUL supervision; training progress subsequently reweights graph edges (Eq.2) and defines distributed healthy/failure reference regions (Eqs.5–7). The HI represents a relative graph-geodesic position between these regions. An online level–velocity filter enforces bounded, non-decreasing HI (Eqs. 8–13). Two experimental findings support the value of this construction. First, A8 demonstrates portability across five backends, not invariance. Second, on the same UMAP representation, HAML features outperform UMAP3 under four decoder families on FD001/FD002 (Fig.G1). The closest simple alternative, single-medoid endpoints (A3), cuts HI–progress correlation from 0.872 to 0.531 (FD003) and from 0.857 to 0.482 (FD004). We will add a comparison with STEP [1], whose self-supervised temporal contrast shapes the embedding itself rather than adding a progress-supervised graph layer afterward, and restrict the novelty claim to this post-representation construction.

3. Fairness and 49-dimensional readout (Q2). We agree that inputs differ. Two controls address this. Table G1 gives baselines descriptor-matched 50×9850 \times 98 inputs: HAML still leads on FD001 RMSE/S-score and on FD003/FD004 P95 Late Error, while several baselines gain on FD002. Fig. G1 (fixed decoder): HAML beats UMAP3 on FD001/FD002, but raw descriptors win with BRF/XGBoost. Baseline hyperparameters follow the original papers and the survey protocol (App.D). We will state the remaining input and tuning differences explicitly. The 49 features retain HI history, velocity, and geometric context; scalar HI alone is insufficient for RUL (A7). G7 assigns 82–98% of BRF importance to HI, graph, and velocity features, but correlated-feature importance is not causal. These controls do not eliminate input-design or tuning confounds; we will clarify this.

4. Geometry versus smoothness (Q3). The decoder relies on the health construction: HI-derived features and the graph coordinate account for 82–98% of forest importance (Table G7), and HAML improves on the unstructured embedding (Fig.G1). The individual graph components (A1, A2, A4) mainly improve lifecycle ordering, with small and sometimes reversed RUL effects. Removing the filter (A5) improves clean-data accuracy but loses monotonicity. This supports an accuracy-constraint trade-off, not universal gains or isolated velocity effects.

5. Training data and 'causal' (W3–W4). Complete training lifetimes supply graph-progress and RUL targets; neither is needed at inference. Figure G23 reduces graph-progress-labelled engines from 100% to 10% with RMSE 13.13→13.6713.13 \rightarrow 13.67 (FD001) and 15.79→15.8915.79 \rightarrow 15.89 (FD004), but the decoder still uses full RUL labels. This is not censored-RUL learning: incomplete histories need censoring-aware treatment, while regime changes and maintenance may violate the irreversible-state prior. 'Causal' means using only observations through time tt, not causal-effect identification or physical-state recovery. We will emphasise 'online/temporally causal' throughout.

[1] Thil et al. (2026). STEP: Learning Structured Embeddings for Progressive Time Series. NeurIPS.

Reviewer_st8UOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates the representation aspect of degradation modelling and RUL-prediction. Authors present HAML which guides the construction of a 'landmark graph', followed by a HI filter. A downstream regressor estimates the RUL. Experiments show promising results.

Critical strengths

The conceptual separation between representation and downstream supervision is clean, and clearly different from, e.g., supervised HI learning. Representing health as a graph-geodesic position is a strong idea, and evidence is provided in the form of numerous ablation studies to test claims.

The persistent level-velocity filter is a very promising solution. Monotonic HI is typically enforced with a cumulative max which is breaks down in the face of noise spikes. Indeed, the ablation shows clearly that removing the filter entirely slightly destroys monotonicity; whereas the cumulative-max alternative keeps monotonicity but degrades performance.

Critical weaknesses

Empirical Results. The empirical case is competitive and promising, but not compelling. HAML is top-3 in only 8 of 12 C-MAPSS comparisons, is clearly beaten by several baselines. Even the PCA-backend ablation (12.91 RMSE / 322 S-score) beats the default configuration on two of three metrics. Are these differences significant? It is not discussed.

Theoretical Insights. The theoretical justification is narrower than implied in the paper, and assumptions are asserted rather than investigated. For example, the progressive time-series assumption is the foundation of the filtering stage, but the paper treats it as a modeling convenience ("modelling assumption") without analysis of when monoticity breaks; early-life transients, condition switching, multi-mode failures, etc. The reader is regularly ponited towards the appendix for important considerations such as theoretical results. That, and the 71 pages (including much of the theoretical discussion) suggests that a conference format might not suit the authors' goal for the paper, and they may want to consider targetting a journal venue.

Related Work. The paper states a conclusion of "explicitly modelling degradation geometry and its temporal evolution offers a principled alternative to unconstrained sensor-to-RUL regression" but such a conclusion is widely known and shared in the literature. The fact that 3 short paragraphs were enough on Relalted Work shows lack of awareness, or more importantly consideration, of related work. This, especially, as the references given there do not include some of the foundational work in an area which goes back several decades. Consider, e.g., [1] -- and several of the literature within. Many methods (including relatively rudimentary architectures and approaches) tabulated there obtain similar or better results on C-MAPPS but are not mentioned in this paper.

[1] Thil et al., STEP: Learning STructured Embeddings for Progressive Time Series (NeurIPS 26)

Questions for authors
  • On FD003, the PCA backend substantially outperforms the default UMAP backend. Can you provide context/discussion to this?

  • The approach carefully pieces apart spatial and temporal features, HI Filter etc., then finally bundles everything together into a 49d representation for predicting the RUL, Γ\Gamma -- doesn't this say that the earlier decomposition in the pipeline is insufficiently effective/relevant to the RUL?

  • Why the graph approach? It's not clear why it is needed apart from a feature extractor. It's not clear that the temporal edges are crucial to performance. Sure, the discussion around it makes sense, but it doesn't add much, in terms of empirical results, as compared to some relatively simple AE architectures, etc.

5 comments
AuthorsOpenReview ↗

Response to reviewer st8U (part 1)

We thank the reviewer for their constructive comments on predictive performance, theoretical foundations, and related work. In response, we conducted additional statistical analyses and investigated the FD003 performance differences between UMAP and PCA. The results highlight HAML's competitive performance, representation-backend sensitivity, and the importance of degradation-aware health construction. We also clarify the theoretical assumptions and distinguish our contributions from existing approaches. We address each point below.

1. Predictive results and statistical significance (W1). Thanks reviewer for the valid point relating to statistical significance. We conducted two-sided Welch's t-tests on ten seed-level runs, with Holm correction (α=0.05\alpha = 0.05) across ten baselines per dataset/bearing–metric group. Of 120 C-MAPSS comparisons (4 subsets × 3 metrics × 10 baselines), HAML significantly outperforms baselines in 36 cases, underperforms in 21, and shows no significant difference in 63 (Table R1). For PHM12, the corresponding counts are 120, 16 and 194 across 330 comparisons (11 bearings × 3 metrics × 10 baselines). Notably, 16 of the 21 C-MAPSS losses occur on FD003, where one engine contributes 72.1 ± 8.7% of the S-score (Fig. G2). Across 15 benchmark tasks, HAML achieves the joint-best average RMSE rank (4.07) and the best S-score (4.1) and P95 Late Error (3.4) ranks. These findings support competitive, albeit dataset-dependent, performance rather than uniform superiority. Detailed statistical results will be provided in the appendix.

Table R1. Statistical significance of HAML versus ten baselines across four C-MAPSS subsets (Welch's t-tests with Holm correction).

RMSE (↓)

DatasetFATLOGOLSTMGRUFC-STGNNSTGNNTimesNetTimeMixer++PatchTSTiTransformer
FD001NS 0.5275NS 0.4336B 0.0096B 0.0096B 0.0004B 0.0001B 0.0483NS 0.1077B 0.0368NS 0.5469
FD002NS 0.5810NS 1.0000B 0.0037B <0.0001NS 1.0000NS 0.1639W 0.0021NS 0.1998B <0.0001NS 0.0549
FD003NS 0.1082NS 0.1524W <0.0001W <0.0001W 0.0077W 0.0021W 0.0015W <0.0001NS 0.1524W 0.0018
FD004B 0.0142W 0.0017B 0.0199B <0.0001W 0.0005W 0.0491W 0.0149NS 0.9786B 0.0017NS 0.2648

S-score (↓)

DatasetFATLOGOLSTMGRUFC-STGNNSTGNNTimesNetTimeMixer++PatchTSTiTransformer
FD001NS 0.1356B 0.0066NS 0.1356B 0.0084NS 0.1356NS 0.1356NS 0.1356NS 0.1356NS 0.1356NS 0.1356
FD002NS 0.0895NS 1.0000B 0.0047B 0.0004NS 0.5554NS 0.3756NS 1.0000NS 1.0000B 0.0161NS 1.0000
FD003W <0.0001W 0.0451W <0.0001W <0.0001W <0.0001W <0.0001W <0.0001W <0.0001NS 0.3184W <0.0001
FD004B 0.0008NS 0.4371NS 0.2045B 0.0002NS 0.4781NS 1.0000NS 1.0000B 0.0321NS 0.1660NS 0.4371

P95 (↓)

DatasetFATLOGOLSTMGRUFC-STGNNSTGNNTimesNetTimeMixer++PatchTSTiTransformer
FD001NS 1.0000NS 0.6863NS 0.1147NS 1.0000NS 0.2553NS 1.0000NS 1.0000NS 1.0000NS 1.0000NS 0.3379
FD002B 0.0013NS 0.4539B 0.0007B <0.0001NS 0.1938NS 0.4539NS 0.4539NS 0.3785B 0.0005B 0.0013
FD003NS 0.1425B <0.0001NS 0.1452NS 1.0000NS 0.3193NS 1.0000NS 1.0000NS 1.0000NS 0.5682NS 0.4023
FD004B <0.0001B <0.0001B 0.0003B <0.0001B 0.0005B 0.0341NS 0.1737B 0.0012B 0.0160B 0.0025

Two-sided Welch–Holm results, α=0.05\alpha = 0.05. Each matrix entry is result code + Holm-adjusted p-value: B (HAML better), W (HAML worse), or NS (No Significance). The correction is applied over ten baselines within each row. Green indicates HAML significantly better; red indicates HAML significantly worse.

AuthorsOpenReview ↗

Response to reviewer st8U (part 2)

2. PCA vs UMAP on FD003 (Q1). UMAP was pre-specified before testing (App. E) and performs best on FD004. However, with the remaining pipeline fixed, PCA improves FD003 RMSE from 14.31 to 12.91 and S-score from 1069 to 322 (third-best S-score in Table 1), highlighting substantial backend sensitivity without isolating the health layer's contribution to FD003's weaker performance. We report A8 as evidence of portability and do not swap backends post hoc.

We investigated FD003's worst engine (#16) across ten seeds. Under UMAP, this engine contributed 72.1 ± 8.7% of the total S-score, with a predicted RUL of 122.16 cycles versus the actual 56 cycles. Switching to PCA reduced its predicted RUL to 84.46 cycles and its S-score contribution by 95.0%, lowering the overall FD003 S-score. To understand this improvement, we examined the health trajectories and embedding neighbourhoods. PCA increased the terminal HI from 0.072 to 0.162, while the median RUL of the 50 nearest training windows decreased from 255.8 to 125.3 cycles. These results suggest that UMAP places the degrading engine near relatively healthy training states, leading to RUL overestimation, while PCA partially corrects this mismatch. We will add this discussion to the Appendix.

3. Why retain 49 decoder features (Q2). The decomposition is used by the decoder rather than discarded. Of the 49 features, 44 describe causal HI history/dynamics; the remainder are raw graph coordinate, persistent velocity and three latent coordinates. Table G7 attributes 82–98% of BRF importance to the HI-related and graph/velocity families, versus 2–18% to latent coordinates. Feature importances are not causal attributions. On FD001, A7 reduces RMSE from 16.08 (current HI only) to 13.69 (44 HI-temporal features), then 13.13 (full readout). A scalar health level cannot alone express degradation rate or mode. The readout is HI-centred, not HI-only or wholly physically interpretable.

4. Why a graph; are temporal edges needed? (Q3) The graph is not a GNN feature extractor: lifecycle-weighted spatial edges discourage cross-stage shortcuts (Eq. 2), and distances to distributed healthy/failure regions define a graph-geodesic health coordinate (Eqs. 5–7). Its main effect is on HI quality. Replacing graph distances with Euclidean ones (A1) lowers HI–progress correlation, and replacing the regions with single medoids (A3) lowers it by 0.34–0.38 on FD003/FD004 (Table G2). These are in-sample ordering diagnostics, not proof of physical validity or superior RUL accuracy. On the same UMAP embedding, adding the health layer beats the raw embedding with all four decoders (Fig. G1). We agree that temporal edges are not essential (A4) and will present them as an optional prior. A decoder-matched autoencoder comparison would further test graph necessity but is not included in the current experiments.

5. Theory and monotonicity (W2). The theorems are conditional, component-level results: the shortcut bound covers only spatial paths, stability assumes fixed OOS neighbourhoods, and nothing implies end-to-end RUL optimality. Monotonicity is guaranteed by projection, but the underlying irreversible-degradation assumption may not hold under early-life transients, regime switching, maintenance or multiple failure modes, which our run-to-failure benchmarks do not test. We will move the theorem statements and a compact “assumptions and scope” paragraph into the main text and shorten the appendix so that the claims are self-contained in the conference format.

6. Related work and novelty (W3). We agree that the current review is too compressed and that degradation-aware representations predate HAML. STEP (Thil et al.) learns a prototype-anchored latent compass using self-supervised temporal contrast, without intermediate lifecycle-progress labels. HAML instead fits a replaceable representation first, then uses training progress to modify graph traversal costs, define distributed endpoint references, and form an online level–velocity state.

STEP reports strong C-MAPSS results, but differences in experimental protocols preclude direct ranking. In detail, STEP excludes test engines with insufficient sequence lengths or ground-truth RUL exceeding 125 cycles, resulting in only 70 of the 100 test engines being evaluated in our FD001 reproduction. In contrast, HAML evaluates all 100 test engines using causal padding for shorter sequences. We will expand the related-work discussion and clarify HAML's novelty as post-representation degradation geometry and online health-state construction.

Thank you to the authors for having substantially addressed my comments. For example, on the empirical results, a range of tests have been enacted. I appreciate that authors have not tried to over-sell the results. The discussion could add value to the paper. Likewise, for providing discussion on PCA vs UMAP, I was interested to see the authors' views and it could benefit the paper to include this.

There are two outstanding points. I accept the explanation of what the graph does in HAML - and this explanation is satisfying, but there was also the implication of 'is it necessary?' (an essential contribution) which I think is still not addressed, or at least not thoroughly.

Regarding the literature: my criticism is not simply that the paper omitted a recent paper, but a general lack of awareness of a mature research area. In their response the authors seem overly defensive. To clarify: I was not intending to doubt or question the novelty or performance of HAML, but point out that readers could benefit from awareness/involvement of other related work, even if a discussion on such comparisons would have to be nuanced due to differences in experimental evaluation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Reviewers' Feedback and Authors' Response

We thank all three reviewers for their constructive feedback on HAML’s predictive performance, technical contributions, experimental fairness, theoretical assumptions, and health-index (HI) evaluation.

Predictive performance and significance (st8U, Qq3N). We conducted additional statistical analyses using ten-seed results with Welch–Holm tests, confirming competitive but dataset-dependent performance. We further investigated FD003’s limitations, showing that PCA improves performance over UMAP and that concentrated prediction errors are associated with representation-backend sensitivity.

Technical contribution and fairness (st8U, Qq3N). We clarify HAML’s contribution as degradation-aware graph geometry introduced after label-free representation fitting, combined with distributed healthy/failure references and online level–velocity filtering. Descriptor-matched and decoder-matched comparisons further support the framework’s contributions.

HI diagnostics and metric definitions (oSrf). We clarify that the original HI diagnostics use complete training trajectories rather than held-out test assets, specify the evaluation population and aggregation, and correct the monotonicity definition in Eq. D5.

Additional age-only and velocity controls (oSrf). New engine-disjoint comparisons against age-only baselines provide evidence of sensor-informed health representations beyond elapsed time. Matched velocity ablations demonstrate improved training-lifecycle HI diagnostics, although they do not establish statistically significant RUL gains.

Theory, assumptions, and causality (st8U, Qq3N, oSrf). We clarify that “causal” refers to online inference without future observations, not causal-effect identification. We also discuss the imposed monotonicity constraint, reliance on complete training lifecycles, and limitations under maintenance, changing operating conditions, and censored histories.

Overall, these additional experiments and clarifications strengthen the empirical evidence, refine HAML’s technical contributions, and ensure that our claims reflect the demonstrated results.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.